How to train VA-VAE?

[lavinal712]

Can you release the code? Thanks for your works!

[JingfengYao]

The training code for VA-VAE is primarily based on the autoencoder training code from LDM. Implementing it should be relatively straightforward, as described in Section 3 of the paper. We are currently considering the most concise way to release it, such as forking or something else.

[lavinal712]

Thanks! I tried to reproduce your code, but I found that the vf_loss does not converge easily. After training for 1000 steps, the model collapsed, and the output turned into solid-color images. Therefore, I would like to see more details in the code.

[gkakogeorgiou]

Hi! Thanks for the great work! I noticed similar behavior to what @lavinal712 observed. Releasing the VA-VAE training code could help us better understand the process.

[lavinal712]

import torch
import torch.nn as nn

from taming.modules.losses.vqperceptual import *  # TODO: taming dependency yes/no?


class LPIPSWithDiscriminator(nn.Module):
    def __init__(self, disc_start, logvar_init=0.0, kl_weight=1.0, pixelloss_weight=1.0,
                 disc_num_layers=3, disc_in_channels=3, disc_factor=1.0, disc_weight=1.0,
                 perceptual_weight=1.0, use_actnorm=False, disc_conditional=False,
                 disc_loss="hinge", 
                 vf_weight=0.1, adaptive_vf=True, vf_loss_type="combined_v3", distmat_margin=0.25, cos_margin=0.5):

        super().__init__()
        assert disc_loss in ["hinge", "vanilla"]
        self.kl_weight = kl_weight
        self.pixel_weight = pixelloss_weight
        self.perceptual_loss = LPIPS().eval()
        self.perceptual_weight = perceptual_weight
        # output log variance
        self.logvar = nn.Parameter(torch.ones(size=()) * logvar_init)

        self.discriminator = NLayerDiscriminator(input_nc=disc_in_channels,
                                                 n_layers=disc_num_layers,
                                                 use_actnorm=use_actnorm
                                                 ).apply(weights_init)
        self.discriminator_iter_start = disc_start
        self.disc_loss = hinge_d_loss if disc_loss == "hinge" else vanilla_d_loss
        self.disc_factor = disc_factor
        self.discriminator_weight = disc_weight
        self.disc_conditional = disc_conditional
        
        self.vf_weight = vf_weight
        self.adaptive_vf = adaptive_vf
        self.vf_loss_type = vf_loss_type
        self.distmat_margin = distmat_margin
        self.cos_margin = cos_margin

    def calculate_adaptive_weight(self, nll_loss, g_loss, last_layer=None):
        if last_layer is not None:
            nll_grads = torch.autograd.grad(nll_loss, last_layer, retain_graph=True)[0]
            g_grads = torch.autograd.grad(g_loss, last_layer, retain_graph=True)[0]
        else:
            nll_grads = torch.autograd.grad(nll_loss, self.last_layer[0], retain_graph=True)[0]
            g_grads = torch.autograd.grad(g_loss, self.last_layer[0], retain_graph=True)[0]

        d_weight = torch.norm(nll_grads) / (torch.norm(g_grads) + 1e-4)
        d_weight = torch.clamp(d_weight, 0.0, 1e4).detach()
        d_weight = d_weight * self.discriminator_weight
        return d_weight

    def forward(self, inputs, reconstructions, posteriors, z_prime, features, optimizer_idx,
                global_step, last_layer=None, encoder_last_layer=None, cond=None, split="train",
                weights=None):
        rec_loss = torch.abs(inputs.contiguous() - reconstructions.contiguous())
        if self.perceptual_weight > 0:
            p_loss = self.perceptual_loss(inputs.contiguous(), reconstructions.contiguous())
            rec_loss = rec_loss + self.perceptual_weight * p_loss

        nll_loss = rec_loss / torch.exp(self.logvar) + self.logvar
        weighted_nll_loss = nll_loss
        if weights is not None:
            weighted_nll_loss = weights*nll_loss
        weighted_nll_loss = torch.sum(weighted_nll_loss) / weighted_nll_loss.shape[0]
        nll_loss = torch.sum(nll_loss) / nll_loss.shape[0]
        kl_loss = posteriors.kl()
        kl_loss = torch.sum(kl_loss) / kl_loss.shape[0]

        # now the GAN part
        if optimizer_idx == 0:
            # generator update
            if cond is None:
                assert not self.disc_conditional
                logits_fake = self.discriminator(reconstructions.contiguous())
            else:
                assert self.disc_conditional
                logits_fake = self.discriminator(torch.cat((reconstructions.contiguous(), cond), dim=1))
            g_loss = -torch.mean(logits_fake)

            if self.disc_factor > 0.0:
                try:
                    d_weight = self.calculate_adaptive_weight(nll_loss, g_loss, last_layer=last_layer)
                except RuntimeError:
                    assert not self.training
                    d_weight = torch.tensor(0.0)
            else:
                d_weight = torch.tensor(0.0)
                
            if z_prime is not None:
                if self.vf_loss_type == "combined_v3":
                    mcos_loss = nn.functional.cosine_similarity(z_prime, features, dim=-1)
                    mcos_loss = nn.functional.relu(1 - self.cos_margin - mcos_loss)
                    mcos_loss = torch.mean(mcos_loss)
                    
                    z_prime = nn.functional.normalize(z_prime, dim=-1)
                    features = nn.functional.normalize(features, dim=-1)
                    
                    cossim_z_prime = torch.matmul(z_prime, z_prime.transpose(1, 2))
                    cossim_features = torch.matmul(features, features.transpose(1, 2))
                    
                    mdms_loss = torch.abs(cossim_z_prime - cossim_features)
                    mdms_loss = nn.functional.relu(mdms_loss - self.distmat_margin)
                    mdms_loss = torch.mean(mdms_loss)
                    
                    vf_loss = mcos_loss + mdms_loss
                else:
                    raise ValueError(f"Unknown vf_loss_type: {self.vf_loss_type}")
                
                if self.adaptive_vf:
                    try:
                        vf_weight = self.calculate_adaptive_weight(nll_loss, vf_loss, last_layer=encoder_last_layer)
                    except RuntimeError:
                        assert not self.training
                        vf_weight = torch.tensor(1.0)
                    vf_loss = vf_weight * vf_loss
            else:
                vf_loss = torch.tensor(0.0)

            disc_factor = adopt_weight(self.disc_factor, global_step, threshold=self.discriminator_iter_start)
            loss = weighted_nll_loss + self.kl_weight * kl_loss + d_weight * disc_factor * g_loss + self.vf_weight * vf_loss

            log = {"{}/total_loss".format(split): loss.clone().detach().mean(), "{}/logvar".format(split): self.logvar.detach(),
                   "{}/kl_loss".format(split): kl_loss.detach().mean(), "{}/nll_loss".format(split): nll_loss.detach().mean(),
                   "{}/rec_loss".format(split): rec_loss.detach().mean(),
                   "{}/d_weight".format(split): d_weight.detach(),
                   "{}/disc_factor".format(split): torch.tensor(disc_factor),
                   "{}/g_loss".format(split): g_loss.detach().mean(),
                   "{}/vf_loss".format(split): vf_loss.detach().mean(),
                   }
            return loss, log

        if optimizer_idx == 1:
            # second pass for discriminator update
            if cond is None:
                logits_real = self.discriminator(inputs.contiguous().detach())
                logits_fake = self.discriminator(reconstructions.contiguous().detach())
            else:
                logits_real = self.discriminator(torch.cat((inputs.contiguous().detach(), cond), dim=1))
                logits_fake = self.discriminator(torch.cat((reconstructions.contiguous().detach(), cond), dim=1))

            disc_factor = adopt_weight(self.disc_factor, global_step, threshold=self.discriminator_iter_start)
            d_loss = disc_factor * self.disc_loss(logits_real, logits_fake)

            log = {"{}/disc_loss".format(split): d_loss.clone().detach().mean(),
                   "{}/logits_real".format(split): logits_real.detach().mean(),
                   "{}/logits_fake".format(split): logits_fake.detach().mean()
                   }
            return d_loss, log

@gkakogeorgiou @JingfengYao Is there any problem of the vf_loss?

[JingfengYao]

I'm not sure about the shapes of features and z_prime. Judging from your mcos_loss = nn.functional.cosine_similarity(z_prime, features, dim=-1) and z_prime = nn.functional.normalize(z_prime, dim=-1), are you placing the channel dimension in the last dimension?

Here are my implementations:

elif self.vf_loss_type == "combined_v3":
    # we also give a 0.25 margin to image dist mat loss
    z_flat = rearrange(z, 'b c h w -> b c (h w)')
    aux_feature_flat = rearrange(aux_feature, 'b c h w -> b c (h w)')
    z_norm = torch.nn.functional.normalize(z_flat, dim=1)
    aux_feature_norm = torch.nn.functional.normalize(aux_feature_flat, dim=1)
    z_cos_sim = torch.einsum('bci,bcj->bij', z_norm, z_norm)
    aux_feature_cos_sim = torch.einsum('bci,bcj->bij', aux_feature_norm, aux_feature_norm)
    diff = torch.abs(z_cos_sim - aux_feature_cos_sim)
    vf_loss_1 = torch.nn.functional.relu(diff-self.distmat_margin).mean()

    # margin_cos
    vf_loss_2 = torch.nn.functional.relu(1 - self.cos_margin - torch.nn.functional.cosine_similarity(aux_feature, z)).mean()
    vf_loss = vf_loss_1 + vf_loss_2

By the way, here are 2 small issues:

1. We have not imposed an upper limit on the adaptive weighting for the vf loss, hence the upper bound value in d_weight = torch.clamp(d_weight, 0.0, 1e4).detach() is set to a relatively large number with d_weight = torch.clamp(d_weight, 0.0, 1e8).detach() specifically for vf.
2. https://github.com/CompVis/latent-diffusion/blob/main/ldm/models/autoencoder.py#L386 There is a particular aspect in the LDM code that requires attention. Please ensure that the newly added projection layer for Z is correctly incorporated into the optimizer.

[lavinal712]

@JingfengYao Thanks for your clarification.
The shape of z_prime is correct. I intentionally placed the channels dimension in the last dimension to align with the dimensionality of DINOv2. The code snippet is as follows:

z = posterior.sample()
z_prime = z.flatten(2).transpose(1, 2)
z_prime = self.proj(z_prime)

This ensures compatibility with DINOv2's expected input format.

And thank you for pointing this out! I agree that the connection layer should be included in the optimizer. I will update the code to ensure that the connection layer's parameters are properly added to the optimizer for training.

[lavinal712]

Well, the model collapses again and I do not know why. Here is the code: lavinal712/VA-VAE

[JingfengYao]

Could you please provide your tensorboard logs?

[lavinal712]

The kl loss is extremely high.

[JingfengYao]

May I ask how many GPUs you use for training and your starting command? @lavinal712

[JingfengYao]

Seems I found the possible reason. Here are my reproductions:

1. I clone your codes and run it with one GPU (I am not sure your starting command. In VA-VAE implementations, I modify to use torchrun for multi-gpu training)

   I start with

   python main.py --base configs/autoencoder/vavae_f16d32.yaml --train --gpu 0,
   

   and get a similar problem with you. My KL loss is also extremely high. That is because of this:

   if opt.scale_lr:
       model.learning_rate = accumulate_grad_batches * ngpu * bs * base_lr
       print(
           "Setting learning rate to {:.2e} = {} (accumulate_grad_batches) * {} (num_gpus) * {} (batchsize) * {:.2e} (base_lr)".format(
               model.learning_rate, accumulate_grad_batches, ngpu, bs, base_lr))
   

   in my setting, lr is automatically set to 2.4e-3, which means a much larger learning rate and smaller batch size than our real training. We use fixed 1e-4 lr and 256 or 512 batch size for training of VA-VAE reported in paper.

2. Then I modify the command as following and log vf_loss before adaptive weighting:

   python main.py --base configs/autoencoder/vavae_f16d32.yaml --train --gpu 0, --scale_lr false
   

   The training becomes more stable:
   

Hope this helps.

[lavinal712]

Thanks! Is it normal for the KL loss to gradually increase?

[lavinal712]

May I ask how many GPUs you use for training and your starting command? 请问您训练时使用了多少 GPU 以及您的启动命令是什么？@lavinal712

4 GPUs,

python main.py --b configs/autoencoder/vavae_f16d32.yaml --t --gpu 0,1,2,3

[lavinal712]

The batch size I used for training the VAE is too small. How many GPUs did you use for training in your paper? Are there any methods to increase the batch size?

[JingfengYao]

Yes, the weight of the KL loss is relatively small among various losses. A tremendous KL loss might impact the generation performance. We utilized 32/64 GPUs to train the VA-VAE. Perhaps you could experiment with mixed precision training, checkpointing, and gradient accumulation (which seems to have been already employed) to increase the batch size.

[lavinal712]

Thank you for your guidance.